Tutorials
In-depth guides for advanced SDK workflows.
| Tutorial | What you'll do |
|---|---|
| Video Sources | Connect USB cameras, RTSP streams, and video files as pipeline inputs. |
| Run Inference in Python (Low level API) | Build a Python application that uses the low level axelera-runtime API to read inference results frame by frame. |
| Run Inference in Python (High Level API) | Build a Python application that reads inference results frame by frame using a high performance gstreamer optimised pipeline. |
| Measure Accuracy | Run a model against a validation dataset and get mAP or top-1 accuracy scores. |
| Deploy Custom Weights | Use your own trained weights with a Model Zoo architecture, through a YAML pipeline and deploy.py. |
| Cascaded Pipelines | Chain two models together: run a classifier on regions detected by a detector, through a YAML pipeline. For the Pipeline Builder Python API, see the Cascade Tutorial. |
| Custom Model Architecture | Deploy a completely new model architecture with a custom decoder and dataset adapter. |
| Compile and Run an LLM | Quantize, compile and run HuggingFace causal-LMs (Llama, Qwen3) on Metis. |
| Pipeline Builder, step 1 Pipeline Builder Quickstart | Compile a YOLO model and run your first Python pipeline on a Metis device. |
| Pipeline Builder, step 2 Model Compilation | Compile an Ultralytics, ONNX or PyTorch model to an .axm for the Pipeline Builder. |
| Pipeline Builder, step 3 Cascade Tutorial | Run a second model on each detection of the first, then trim, track and nest cascades. |
| Pipeline Builder, step 4 Coordinate System Tutorial | Map boxes, keypoints and rotated boxes from any cascade level back to the original frame. |
| Pipeline Builder, step 5 Profiling | Read a pipeline's per-operator report and find what limits throughput. |
| AxScript, step 1 Compile and Run an AxScript Model | Build one of the AxScript cards this SDK ships, run it on a Metis device, and read what it prints, before changing anything. |
| AxScript, step 2 Serve an AxScript LLM through HuggingFace, or Your Own Front End | Load a built text or vision-language artifact with the Auto classes and your own sampling parameters, or drive the same engine from a front end of your own. |
| AxScript, step 3 Build an AxScript Model from Your Own Weights | Point a vision, text or vision-language card at your fine-tune, your output head, and your calibration data. |
The Pipeline Builder reference (pipeline overview, model formats, postambles, API) is at Pipeline Builder.